Generate Block not working properly

[nikghost17]

I am creating a rag agent

so basically I am creating a chatbot for finantial document
as you can see I am performing different steps so that retrieval will be spot on

Ideal behaviour -->>>

from retrieval_1 (from loop) and retrieval_0 (after the keywords) I am getting some chunks of data (piece of data from the pdfs) for a specific user query (user asked question)
that chunks I want to pass to my llm (named 'Relevance-Score-Calculator' (the 'Relevance-Sco...'), llm is open-AI gpt 4o) for some operation

problem in this -->>>

when I pass these chunks there are a few duplicates
which if I pass to llm it exceeds the limit of my llm

now for that I have added another generate block name unique-chunks-passer
it is supposed to pass unique chunks futher

the prompt-->>

now dispite this the generated output is the answer to the user question not the unique chunks array

##INPUT->>

##OUTPUT

[nikghost17]

@ZhenhangTung
@yongtenglei

please need help !!

[dosubot[bot]]

It looks like the "unique-chunks-passer" block is outputting an answer instead of a JSON array of unique chunks because the LLM is not following the prompt instructions strictly enough. In RAGFlow, the generate block just sends your prompt and input to the LLM and returns whatever the model outputs—there’s no backend enforcement of output format or post-processing to guarantee a JSON array. If the prompt isn’t clear or the input isn’t formatted as expected, the model may default to answering the user question instead of extracting unique chunks as you want.

To fix this, try these steps:

• Make your prompt as explicit as possible: for example, “Extract only the unique, valid chunks (enclosed in curly braces) from the list below. Output ONLY a JSON array of these unique chunks, with no explanation or extra text. Example output: ["chunk1", "chunk2"]”.
• Add a clear example input and output in your prompt.
• If your input chunks are delimited, make sure you use backticks around your delimiter (e.g., ##) instead of single quotes, and reduce the chunk token size if needed to avoid merging chunks together [source].
• If your model supports it, add the directive //no_thinking to your prompt to suppress reasoning and enforce stricter output [source].
• If you still get inconsistent output, consider adding a post-processing step to validate and parse the LLM output as a JSON array.

If you can share your exact prompt and a sample input, I can help review and suggest improvements tailored to your workflow.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[nikghost17]

@dosu
Yes tell me what exactly you want I am ready to give you

[dosubot[bot]]

To help you fix the "unique-chunks-passer" block, could you please share three things: (1) the exact prompt text you're using in that block, (2) a sample of the actual input data (the list of chunks as it appears when passed to the block), and (3) a sample of the actual output you're getting (the incorrect answer instead of the expected JSON array)? With these, I can review and suggest specific changes to your prompt or workflow to get the output you want. This info is key because small differences in prompt wording or input formatting can make a big difference in how the LLM responds.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[ZhenhangTung]

@nikghost17 kinda lost in your question. are you asking why there are duplicates and the content exceeds the context window?